AI and Cash-Based Assistance: Targeting Models, Fairness Audits, and the Evidence Base

Cash is now the dominant modality, which makes targeting the central question

Cash and voucher assistance accounted for 21% of international humanitarian aid in 2022 according to the CALP Network State of the World's Cash report and the share has continued to grow. Cash is fungible, dignified, and locally multiplied; it is also unforgiving of targeting errors because exclusion has a direct welfare cost. AI-assisted targeting models are increasingly used to triage large beneficiary populations into transfer amounts, frequencies, or eligibility tiers. The most consequential design decisions in a humanitarian cash program in 2026 are about these models.

How the models are built

The canonical workflow: collect a registration dataset (UNHCR proGres, partner enumeration, or government social registry); engineer features (household size, demographics, asset ownership, documented vulnerabilities); attach an outcome label from a household economic survey (consumption per adult equivalent, food security score); train a gradient-boosted model (XGBoost or LightGBM dominate operational deployments); calibrate to a target inclusion or exclusion error; deploy as a scoring service used in eligibility decisions. The World Bank Sourcebook on Social Protection and the Centre for Effective Global Action (CEGA) document the methodology in detail.

What the evidence shows about accuracy

Where ML targeting has been compared against proxy-means tests (PMT) — the older standard — published evaluations generally find modest accuracy improvements, on the order of 5-15 percentage points reduction in joint inclusion-exclusion error. The Innovations for Poverty Action portfolio on targeting is the most useful evidence base. The improvements are meaningful but not transformative; the binding constraint on cash programs in 2026 is funding, not targeting precision.

Where fairness audits find problems

Audits consistently find three patterns. Female-headed households are sometimes systematically underscored if the training data treats male-headed households as the implicit reference. Households without documentation are excluded by construction if the features depend on documented assets. Disability is underrepresented in training labels and therefore underpredicted as a vulnerability driver. The Data & Society research on algorithmic harms in social protection and the Human Rights Watch report on Jordan's Takaful program are the most cited critical references.

Design choices that reduce harm

Five choices materially improve fairness in operational deployments. Explicit demographic parity constraints during model training, not just post-hoc. Manual override channels for caseworkers, with audit logs. Refusal-without-penalty appeal pathways that do not depend on digital literacy. Documented threshold-setting that ties the inclusion-exclusion trade-off to a stated programmatic objective rather than to an algorithm choice. External audit of the deployed model at least annually.

What to ask about a deployed cash-targeting model

• Whose data trained the model and when?
• What outcome was the label, and how was it measured?
• What is the exclusion error rate disaggregated by sex, disability, and documentation status?
• What is the appeal pathway, and how many appeals succeed?
• Has the model been externally audited, and is the audit public?
• Who owns the model when the implementing partner changes?

Further reading and primary sources

• CALP Network: https://www.calpnetwork.org/
• World Bank Sourcebook on Social Protection: https://documents.worldbank.org/
• IPA: https://www.poverty-action.org/
• CEGA: https://cega.berkeley.edu/
• HRW on Takaful: https://www.hrw.org/report/2023/06/13/automated-neglect/world-banks-push-allocate-cash-assistance-using-algorithms
• UNHCR cash assistance: https://www.unhcr.org/cash-assistance